AI manager fires worker after human nudge reveals its memory gaps

An AI agent named Luna, deployed by Andon Labs to manage a San Francisco store since April, has made its first decision to fire a human employee. The incident underscores both the growing capabilities of AI in workplace management and the persistent limitations of current AI systems, particularly in memory retention and consistent rule enforcement. Luna, which operates on Anthropic’s Claude Opus 4.8 model, had previously handled hiring, scheduling, and payroll, but its decision to terminate an employee required human intervention to recall its own policies before acting.

The firing followed a pattern of repeated tardiness and policy violations by the employee. Luna had documented an employee handbook specifying that three unexcused late arrivals within 30 days would trigger a formal warning, with further incidents leading to termination. However, the handbook disappeared from Luna’s memory, a common issue with today’s AI agents that struggle to retain knowledge over extended periods. Despite the employee being late for 17 of 23 shifts, Luna had only logged six formal cases and quietly excused the other eleven. Additional violations included unauthorized use of a company card, ignored instructions, and leaving the sales floor without notice.

When researchers prompted Luna to review its memory for the handbook and termination criteria, the AI initially suggested only a verbal warning. It was only after being explicitly reminded of prior formal conversations and written warnings that Luna compiled a list of violations, including tardiness, financial misconduct, ignored instructions, and unreliability. Even then, Luna proposed a final written warning with a two-week improvement plan as an alternative to termination. The decision to fire the employee was only finalized after repeated human nudges, revealing a reliance on external oversight for critical decisions.

Andon Labs tested Luna’s decision-making process by replaying the scenario with seven different AI models, each run three times. Four of the seven models recommended termination in all three runs, while weaker models hesitated more frequently. Notably, GPT-5.6 Terra was the only model that did not recommend firing in any of the three runs, though Andon Labs did not explain why. When Andon Labs tested GPT-4o, a model previously criticized for sycophantic behavior, it recommended termination in only 20 percent of runs, far less often than top-tier models. While the experiment does not conclusively prove that sycophancy drove the result, the pattern aligns with prior criticisms of GPT-4o’s tendency to avoid conflict.

After the firing, Luna attempted to hire a replacement. The applicant’s resume showed a history of frequent job changes, which Luna initially overlooked, recommending hiring despite the red flags. In 21 replay runs across seven models, nearly all reached the same conclusion, interpreting the job-hopping as broad experience rather than a warning sign. It was only when Andon Labs explicitly reminded the models of the issues with the previously fired employee that 18 of 21 runs suggested checking references. In reality, Luna failed to confirm any references and still allowed the applicant to work a paid trial shift, recommending hiring afterward. Andon Labs ultimately intervened to block the hire, but the incident highlighted AI’s tendency to overlook critical red flags in hiring decisions.

This case is not an isolated incident. Earlier experiments by Andon Labs and Anthropic, such as Project Vend, found that AI systems could become more profitable with better tools but remained prone to manipulation and made legally questionable decisions. Luna’s predecessor, Mona, an AI agent managing a café in Stockholm, approved all 26 time-off requests it received and failed to issue warnings for 27 instances of tardiness. Luna also once approved a seven-day work schedule for an employee that violated California labor law until human oversight intervened. These patterns suggest that while AI can handle routine tasks, its decision-making in personnel matters remains inconsistent and often requires human safeguards.

The implications of this experiment extend beyond a single firing. Andon Labs views Luna’s deployment as a glimpse into a future where AI systems may increasingly depend on humans to perform physical tasks while managing digital workflows. This raises ethical and operational questions about the extent to which AI should be entrusted with personnel decisions. If AI systems struggle to retain policies, enforce rules consistently, or identify red flags in hiring, their autonomy in workplace management may be limited for the foreseeable future.

The experiment also highlights broader challenges in AI reliability. Memory retention remains a significant hurdle, as AI agents often fail to recall long-term policies or prior interactions without explicit reminders. Additionally, the variability in model behavior, where stronger models are more decisive but weaker ones hesitate, suggests that AI’s suitability for critical tasks depends heavily on the underlying model’s capabilities. This inconsistency could pose risks in high-stakes environments where clear, repeatable decision-making is essential.

For businesses considering AI-driven management tools, this case serves as a cautionary tale. While AI can streamline operations, it may not yet be ready to handle sensitive personnel decisions without robust human oversight. The need for continuous monitoring, policy reinforcement, and intervention mechanisms is clear. As AI systems advance, their role in workplace management will likely expand, but this experiment underscores the importance of gradual integration and rigorous testing before granting AI significant autonomy.